Make a new picture from a photo you already have: give each reference a clear role, then check which details carried over and which changed.
Try an example
A desk lamp photo, and the result generated from it with the request “Place the lamp from the reference on an uncluttered dark desk with soft light from the left.” The lamp is guided by the reference; the desk and light are generated.
The lamp’s overall body and shade shape follow the reference, and the new desk keeps the lamp as the prominent object. The reference guides the model’s interpretation and does not lock any area, so carried over means similar, not identical.
The workspace is new: a darker, uncluttered desk with soft light from the left. Framing and shadows are generated, and the lamp itself can show small differences that come from reinterpreting the reference.
Compare the stem, switch position, shade opening and visible cable with the real lamp. A convincing concept can still misrepresent a small detail. If the job is to document the real lamp, use the original photo or a background tool instead.
The result depends on what the reference carries, so each kind of reference needs a different request and a different review.
4:3 · Product setting
Start from a clear photo where the product’s surfaces and edges are visible, then state the setting, the placement and the light. A quiet work surface with soft side light is enough. Compare the product with the source before judging the background, because a convincing scene can still misrepresent a small port or seam. Treat the result as concept art, not a catalog record.
4:5 · Portrait treatment
A portrait reference can guide an illustrated avatar or an editorial concept, but a generated face may differ from the source. Use a clear photo and ask for one treatment, such as softer window light or a graphic palette. Compare expression, hair, clothing and silhouette with the source; a recognizable interpretation is not an unchanged photograph.
1:1 · Illustration set
Choose a reference that makes the treatment visible, such as line weight, texture or palette, and describe the new subject separately so the prompt does not repeat the original scene. A paper-style plant can guide a desk lamp. Review the images as a group; line weight and lighting can drift enough to make one image feel out of place.
Three steps in the creation card above.
Pick an image that shows the feature that matters clearly, such as the front panel of a product or the full pose of a figure. Avoid heavy filters, interface screenshots and crowded collages, and keep the original for review. The generator accepts up to 14 reference images, subject to the model; use only images you are entitled to use.
In the prompt, give every reference one job, for example: use the first image for the lamp, the second for the color palette. Say which wins if they disagree. Add another reference only when it answers a question the others cannot.
Choose the aspect ratio for the placement, then model, resolution and one, two or four outputs. One Nano Banana HD image costs 23 credits and one Nano Banana Pro HD image costs 45, shown on Generate. Compare the result with the original subject first, then judge the crop at real size.
Four ways a reference image shapes the result, and what to check in each.
01
Image to image AI starts from a picture you already have. Say whether a reference supplies the subject, the setting, the colors or the style, so the request does not blur them together. With several references, name the job of each one in the prompt before you generate.
02
A product photo can move to a quiet surface with new light while the object stays recognizable. The reference guides the subject, and the surroundings are generated from your words. Compare the result with the source for shape, labels and proportions before you use it.
03
A portrait can be repainted as an oil painting, redrawn in ink or turned into a cartoon while the composition follows the original. The reference sets pose and framing, and the prompt sets the treatment. Check faces and hands, because a treatment can change details the reference showed clearly.
04
The result does not have to match the ratio of the reference. Choose the aspect ratio and resolution for where the new image will be used, and the quote on Generate updates with every change. PNG or JPEG sets the file you download.
Pick the page that matches the change you want to make next.
Image to image AI generates a new image using one or more existing images with a written request. The references guide visible information such as a subject, layout, or treatment. Potto AI uses this as reference-guided generation. The output can reinterpret details, so it should not be treated as an exact copy or a guaranteed identity-preserving edit.
A reference is visual material supplied to the request; a style selector is a specific interface control. Explain which part of the reference should guide the result. This workflow has model and output controls, but an example style shown in the page is not automatically another available menu option.
They can provide useful information, but consistency still needs comparison. Use clear, complementary references and identify stable features such as silhouette, colors, and distinctive objects. Reject contradictory or inaccurate results. A collection of references does not create an identity lock or a verified product specification.
One that shows the feature that matters clearly. If the task depends on a product’s front panel, a distant lifestyle shot carries too little detail; if it depends on a pose, a crop that cuts off the hands omits it. Avoid heavy filters, screenshots with interface overlays and crowded collages, and keep the original available for review. Enlarging adds pixels but not information that was never visible. Use only images you are entitled to use.
Write each role into the prompt in ordinary language: which image supplies the subject, which the composition, which the lighting, treatment or color direction, and which wins if they conflict. The generator takes up to 14 reference images, subject to the selected model, but that limit is a ceiling, not a target. Afterward, trace important details back to their intended source.
Use the AI image editor when you have a selected picture and can describe one specific change; a written instruction is not a precise mask, so check the rest of the image. Use a dedicated tool when the operation itself is the task, such as removing a background, expanding a canvas or enlarging an image. Use a template for a fixed creative treatment. Use image to image AI when a source should guide a new scene or interpretation.
You can upload a reference image, write the instruction and set the output before signing in. Generate opens the sign-in dialog, and the reference and instruction stay in place while you sign in.
A failed request shows the reason and a retry option, and its credits are released. If the reference did not upload, add a usable JPG, PNG or WebP file. Stop a running request with its cancel control; a new attempt on a finished result is charged again.
The cost follows the model, resolution and number of images, not the number of references: 45 credits per image with Nano Banana Pro at HD or 2K, 23 with Nano Banana at HD. Generate shows the total before you confirm.
Add a reference, say what it should contribute, and compare the result with the source before you download.